If you find yourself typing “be brief,” “skip the bullet points,” or “I work in an accounting firm” at the start of every single chat, custom instructions exist precisely to end that repetition. ChatGPT, Gemini and Claude each provide a settings field where you describe once who you are and how you want answers delivered, and the assistant applies that context across all conversations. The catch is that this text is not a private drawer: it lives in your account, and depending on the service and your data controls, it may feed model training. Here is where each option sits, what actually improves output, and what should stay out of the box.
- Three different menus, one idea: OpenAI, Google and Anthropic all bury the feature in account settings, but the names, limits and extra layers differ significantly.
- Specificity beats politeness: Instructions that define context, format, language and what you do not want outperform vague requests to “be clear.”
- Privacy is the real risk: OpenAI states custom instruction data is used to improve models unless you opt out, while Google retains reviewed conversations for up to three years.
Where each chatbot hides the setting
Custom instructions should not be confused with memory, which is what the model infers or stores from your conversations. Instructions are what you dictate explicitly. All three tools apply them globally rather than to a single thread, but the paths diverge.
ChatGPT: instructions plus personality layers
On the web and desktop, OpenAI routes users through Settings > Personalization, where you confirm that Enable customization is switched on and then write into the Custom Instructions field. On mobile the path is Settings > Customize ChatGPT. The feature is available on every plan, but the character limit scales with subscription tier: as of November 2026 it stands at 1,500 characters for Free and Go, rising to 5,000 for Plus, Pro, Business, Enterprise and Education.
ChatGPT adds two structured layers beyond free text. The first is personality, with presets including Default, Professional, Friendly, Candid, Quirky, Efficient and Cynical. The second is Characteristics, which dial warmth, enthusiasm, use of headings and lists, and emoji density up or down. Both operate alongside instructions and memory rather than replacing them.
Gemini: personal intelligence, with regional naming
On gemini.google.com, Google places the feature under Settings and help > Personal intelligence > Instructions for Gemini. You press Add and enter each instruction individually; from the same screen you can edit, delete or disable them all at once. Google notes that all users can add instructions, though its requirements list a personal Google account rather than a work, school or supervised one. In some regions the feature appears under the label “Saved info.”
Claude: global and per-project scopes
Anthropic’s help centre locates the field under Settings > Your Claude > Instructions, or Account > Instructions for Claude in the older layout. Claude additionally separates project instructions, which apply only to chats inside a specific project. Anthropic’s documentation is inconsistent here: one section says projects are available on free accounts with a cap of five, while another presents project instructions as a paid-plan feature. Verify against your own account before relying on it.
What to actually write in the field
The practical rule is to write what you would tell someone starting work with you on Monday: context, formatting preferences and boundaries. Anthropic recommends including preferred methods, recurring terminology, typical scenarios and communication guidelines, while Google offers examples such as opening every response with a short summary.
Instructions that outperform a generic “be clear” include:
- Professional context without personal data: “I work in customer service at an insurance company. When I ask for text, it is for clients in Spain.”
- Skill level: “I code in Python at a basic level; explain anything non-obvious, but do not explain what a variable is.”
- Format: “Answer in two sentences first, then expand. No tables unless I ask.”
- Language and register: “Write in British English, informal. If a technical term is in English, keep it and explain it once.”
- What you do not want: “Do not agree with me by default. If you spot an error in my reasoning, flag it before continuing.”
An instruction demanding explicit pushback is arguably the highest-value line you can add, because it counteracts the sycophancy these systems tend toward. For narrower tasks, the general prompt-writing techniques published by OpenAI, Anthropic and Google are the better reference.
Two caveats from the fine print. Google warns that instructions telling Gemini to forget or avoid a topic “do not always work perfectly,” and that deleting the relevant conversations is the reliable fix. And if you need different behaviour for different tasks, do not cram everything into the global field: projects and custom assistants exist for exactly that purpose.
The privacy question: what should never go in
The baseline is simple: do not put anything in custom instructions that you would not send in a message to the company running the chatbot. This is account data, it is stored, and depending on the service and your settings it may be used to improve models.
OpenAI is explicit that information in custom instructions is also used to improve model performance unless you disable that option in your data controls. It adds that instructions are not shown to anyone viewing a shared chat link, that they are included in data exports, and that deleting your account removes them within 30 days. OpenAI also warns that third-party plug-ins may receive relevant instruction details.
Google lists Instructions for Gemini among the data it collects and states they are retained until you delete them. The same privacy documentation notes that human reviewers examine a portion of conversations, that reviewed conversations are kept for up to three years without being linked to the account, and asks users not to enter confidential information they would not want a reviewer to see.
Given those terms, keep the following out of the field:
- National ID numbers, social security numbers, IBANs or any account or card number.
- Passwords, access codes or answers to security questions.
- Medical diagnoses, medication or family health data.
- Company name, exact job title and city if together they identify you.
- Third-party data: children, partners, clients or colleagues by full name.
The alternative is to describe without identifying. “I work at a small business in the healthcare sector” delivers the same useful context as naming the clinic.
When to revisit your setup
Review the field whenever responses stop fitting, and at minimum when you change jobs or your primary use case. Instructions written for exam preparation become noise once you are using the chatbot for cooking.
Disabling is straightforward: switch off Enable customization in ChatGPT, toggle off Instructions for Gemini in Google’s tool, or edit the field in Claude. A simple test is to ask the same question with instructions on and off; if the output is identical, the words are surplus.
Custom instructions remain one of the most underrated features in consumer AI and, simultaneously, the one that most invites over-sharing. The sensible balance is to treat them as a style guide rather than a personal profile: how you want to be addressed, in what format and at what level, with nothing that reveals who you are.